Clustering & Customer Segmentation Explained: K-Means Process
After Classification Models Explained for Analysts, the next interview jump is from labelled classes to unlabelled customer groups. Clustering answers the practical customer segmentation question: how to group customers when there are no predefined labels. For analysts, the interview focus is usually the K-Means process, the choice of k, and when to choose K-Means versus Hierarchical, DBSCAN, or Gaussian Mixture.
- Clustering is an unsupervised technique that groups unlabelled data into natural clusters.
- K-Means Clustering is an unsupervised algorithm partitioning n points into k clusters by minimising within-cluster sum of squares.
- The K-Means process is: choose k, initialise centroids, assign points to nearest centroid, update centroids, and repeat until convergence.
- K-Means is used for spherical, similar-size clusters and is fast on large data, but it assumes spherical clusters and is sensitive to outliers.
- Hierarchical is useful when the number of clusters is unknown and dendrogram visualisation is needed.
- DBSCAN is useful for clusters of arbitrary shape and handling noise/outliers, while Gaussian Mixture is useful for soft/probabilistic cluster assignment and elliptical clusters.
- Customer segmentation can use Recency, Frequency, Monetary (RFM) scoring, K-Means clustering, and demographic segmentation, with segment size and Avg Customer Lifetime Value (CLV) per segment as key metrics.
Big Picture: Clustering as Customer Segmentation
Clustering is an unsupervised technique: group unlabelled data into natural clusters. In customer analytics, Customer Segmentation can use RFM scoring, K-Means clustering, and demographic segmentation; the key metrics are segment size and Avg Customer Lifetime Value (CLV) per segment.
For the question "How to group customers?", the framing is: no predefined labels, then normalise, cluster, profile, and produce labelled segments.
K-Means Clustering Process
Choosing the Right Clustering Algorithm
The algorithm choice depends on the cluster shape, the number of clusters, the required output, and the limitation the analyst can accept. K-Means is only one option in the clustering toolkit.
Choosing k and Reading the Output
K-Means requires k, the number of clusters. The process starts by choosing k using Elbow/Silhouette, and Silhouette Score is a clustering metric: (-1 to 1) - measures cluster cohesion and separation.
The output is a hard cluster assignment for each point. This is useful when the analyst needs labelled segments, but the limitation is that K-Means assumes spherical clusters and is sensitive to outliers.
Customer Segmentation in Practice
Customer segmentation uses RFM scoring, K-Means clustering, and demographic segmentation. RFM means Recency (days since last purchase), Frequency (# orders), Monetary (₹ LTV).
The key metric is segment size and Avg CLV per segment; the tool is Python (sklearn), SQL; the Indian example is Myntra segmenting fashion shoppers into 6 RFM tiers.
CLV = Average Order Value (AOV) × Purchase Frequency × Customer Lifespan. Simple CLV - useful for first-pass segmentation decisions.
When Analysts Should Pick K-Means Versus Other Algorithms
Choose K-Means when clusters are spherical, similar-size, and speed on large data matters. Choose Hierarchical when the number of clusters is unknown and dendrogram visualisation is needed.
Choose DBSCAN when the problem involves clusters of arbitrary shape and handling noise/outliers. Choose Gaussian Mixture when the output needs soft/probabilistic cluster assignment and the clusters are elliptical.
Structuring a Clustering & Customer Segmentation Explained Interview Answer
"How would you group customers when there are no predefined labels, and when would you choose K-Means over Hierarchical, DBSCAN, or Gaussian Mixture?"
Do not jump straight to K-Means as the default answer. Start from the business question, confirm that there are no predefined labels, then normalise, cluster, profile, and produce labelled segments.
The most frequent error is choosing K arbitrarily. It costs points because K-Means specifically depends on k (number of clusters), and the process expects Elbow/Silhouette rather than a random choice.
Conclusion
Clustering is the analyst’s way to group unlabelled customer data into natural segments, with K-Means as a fast option for spherical, similar-size clusters. The strongest interview answer explains the K-Means process, compares algorithm choices, and avoids choosing k arbitrarily.